acceptodds
Under review as a conference paper at ICLR 2027

Pathwise Relational Conformer Tensors for Molecular Ensemble Learning

Abstract

Molecular conformer ensembles are often modeled as sets of three-dimensional structures, yet they possess substantially more structure than generic sets: all conformers share the same molecular graph, inducing exact correspondence among atoms, bonds, angles, and torsions. We argue that these aligned local structures provide natural sites for relational computation before ensemble aggregation. We introduce PaSTNet, a molecular ensemble architecture built around the Pathwise State Tensor (PaST) operator. For each molecular path, PaST lifts conformer path states into channel-valued symmetric and antisymmetric pair tensors, contextualizes these relations using permutation-equivariant tensor operations, and contracts them back into conformer updates. We characterize PaST's linear core and show that its structural maps form complete families of homogeneous permutation-equivariant linear operators within the symmetric and antisymmetric pair branches at fixed conformer cardinality. To center each block's ensemble response relative to a singleton reference, PaST further employs a singleton-centered residual that subtracts the response of the same relational operator re-evaluated on the corresponding singleton input, making each PaST block exactly the identity when only one conformer is available. PaSTNet alternates these relational updates with geometry-conditioned propagation across molecular paths, allowing cross-conformer relational signals induced by conformational differences to propagate through the atom–bond–angle–torsion hierarchy before conformer pooling. Together with reversal-invariant path encoding and invariant geometric descriptors, the resulting model is invariant to a common permutation of conformers, consistent atom relabeling, and independent proper rigid motions of each conformer. Across six molecular property tasks, PaSTNet achieves the best evaluated mean score on five. Identity, attention, pair-local, and no-subtraction controls favor full PaST in mean performance; attribution remains limited by coupled changes in some controls and uncertainty over three splits. PaSTNet therefore provides a principled framework for reasoning explicitly over aligned local relations in molecular conformer ensembles rather than treating conformers only as representations to be pooled.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
molecular representation learning, conformer ensembles, geometric deep learning, molecular property prediction, equivariant neural networks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_AzHqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a molecular conformer ensemble learning framework. It models pairwise relationships between corresponding local structures across conformers. The method also introduces symmetric and antisymmetric relational tensors, together with geometry-aware message passing and also a singleton-centered residual. I find the idea clear and interesting, but have some concerns remained about the method novelty and experimental evaluation.

Critical strengths
  1. The paper is well organized and the motivation is easy to understand: instead of only pooling conformers, it tries to model aligned local relationships across conformers.
  2. The method is technically detailed, and the theoretical properties of the proposed operator are stated clearly.
  3. The empirical study covers multiple molecular property prediction benchmarks and includes several controls rather than only a main-table comparison.
Critical weaknesses
  1. The idea of explicitly modeling pairwise relationships across conformers is relatively interesting, but the novelty should be better clarified. For example, the tensor operations are largely adapted from existing constructions. Then, the prior work has already considered cross-conformer interactions, which leads that it remains somewhat unclear why retaining explicit pairwise states is especially useful beyond being a different architectural choice.

  2. I also noticed that the Table 2 compares PaST with a parameter-matched attention baseline, but the improvements are relatively small compared to the reported standard deviations. Also, the two variants differ in several aspects beyond the use of explicit pairwise states, so it is hard to pinpoint what actually causes the improvement.

  3. The computational cost is another concern. Since the method explicitly constructs pairwise tensors, its memory and computational costs should scale with the number of conformers. Table 4 shows noticeably higher memory and inference cost than FACET at K = 10, while the experiments only evaluate up to 10 conformers.

  4. I thinks the singleton-centered residual is an interesting design, but its practical benefit is not fully claimed from the current experiments. Althougn Table 3 uses separately trained models for different conformer counts, it does not show how well a fixed checkpoint handles varying numbers of conformers when inferencing.

Questions for authors
  1. What do explicit pairwise states offer over existing methods? Could the authors give a more concrete example to claim the contributions better?

  2. Could the authors provide a more controlled comparison with some attention or relational baselines if possible? Which will help isolate the effectiveness of pairwise-state retention more directly.

  3. Since the computational cost grows quadratically with the number of conformers, I wonder how well would PaSTNet scale beyond K = 10?

  4. Table 3 uses separately trained models for different values of K. What happens if the same trained model is tested with different numbers of conformers? Would singleton subtraction make a difference in this setting?

Reviewer_PGibOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

In this paper the author introduces PaSTNet, a molecular property prediction model that explicitly models relationships between different conformers of the same molecule before combining their information. The key idea is to leverage the shared molecular structure across conformers and enable interactions between them through a new PaST operator. The method is designed to preserve important physical symmetries and is built on top of a SchNet backbone. Experiments on six MoleculeNet datasets show strong performance, achieving the best average results on five datasets. The authors also provide several ablation studies, which generally support the benefit of the proposed components, although they note that the performance differences are not statistically significant.

Critical strengths

I think the core idea of the paper is strong. Rather than pooling conformer representations early, the method explicitly models how aligned conformers differ from one another, making better use of the shared molecular structure across the ensemble. The proposed pairwise interactions between conformers provide a principled way to capture conformational variation instead of simply averaging it away. The symmetric and antisymmetric decomposition is also well motivated and aligns naturally with different types of relationships between conformers.

Another strength is the authors' careful evaluation. Baselines are reimplemented consistently, parameter-matched controls are included. The authors clearly acknowledge where theoretical guarantees apply, where ablation studies have confounding factors, and where statistical significance cannot be established. They also discuss comparability issues with external baselines. This level of transparency makes the empirical evidence easier to trust and the paper easier to assess fairly.

Critical weaknesses

I am not convinced that the additional complexity delivers a clear benefit. The reported gains over strong baselines are generally small and often within the variation across data splits. Moreover, the K = 1 version, where the proposed PaST module effectively does nothing, already achieves highly competitive performance. This suggests that much of the performance gain may come from the backbone and path-based architecture rather than the conformer interaction mechanism itself.

Finally, the paper does not sufficiently connect the model to physically meaningful ensemble behavior. The conformer weighting scheme is not related to Boltzmann populations, and the method is not evaluated on datasets that include conformer energies. At the same time, computational cost grows quadratically with the number of conformers, raising questions about scalability to larger and more realistic conformer ensembles.

Questions for authors

For chiral molecules in ESOL, do enantiomer pairs get the same prediction? Given the signed torsion features I would expect not always, and I am curious how the authors think about this for properties that are identical for enantiomers.

How were the conformers for the MoleculeNet datasets produced and how many per molecule, which force field, what energy window, how were duplicates handled? How much do the results move if a different conformer generator is used?

Reviewer_GzGkOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes PaSTNet, a model for molecular property prediction from conformer ensembles. Its central observation is that all conformers of a molecule share the same graph, so each covalent path (atom, bond, angle, torsion) is exactly aligned across conformers and can serve as a site for cross conformer relational computation. The core operator, PaST, lifts the K conformer states at each path into symmetric and antisymmetric channel valued K by K pair tensors (with relative bond length and angle or torsion differences as extra features), contextualizes them with two permutation equivariant tensor layers built from diagonal, row, trace, and global contractions, and contracts them back into per conformer updates. A singleton centered residual subtracts the same operator evaluated on each conformer alone, so each PaST block is exactly the identity when K = 1. PaST alternates with Geo IH, a geometry conditioned message passing step over the path hierarchy within each conformer (adapted from PCNN), on top of a shared SchNet backbone, followed by pathwise conformer pooling, a topology only refinement sweep, and an MLP head. The authors prove that the branchwise linear mixers are complete families of homogeneous permutation equivariant maps at fixed K (a specialization of Maron et al.), and that the predictor is invariant to common conformer permutation, independent proper rigid motions of each conformer, and consistent atom relabeling. On six MoleculeNet tasks with locally reproduced baselines under random scaffold splits, PaSTNet reports the best mean on five of six. Controls with an identity block, a parameter matched attention block, a pair local variant, and a no subtraction variant favor full PaST in mean score, a conformer cap study shows nonmonotonic behavior, and profiling shows PaSTNet costs several times more than FACET at K = 10. The paper is explicit that it makes no significance claims and that attribution to individual components remains open.

Critical strengths
  1. The representational idea is clear and well motivated. Conformers of one molecule share the molecular graph, so every covalent path is exactly aligned across the ensemble, and PaST uses each path as a site for an explicit channel valued pair tensor with symmetric and antisymmetric branches that is contextualized before contraction. This sits in a sensible middle ground between pooling, which discards pair structure, and standard attention, which reduces each pair to a scalar score immediately. Alternating PaST with Geo IH so that cross conformer signal can propagate through the atom, bond, angle, torsion hierarchy is a natural way to use the path complex view.

  2. The theory is stated carefully and at the right level of modesty. Theorem 1 characterizes the branchwise linear mixers as complete families of homogeneous permutation equivariant maps at fixed K, with an explicit orbit enumeration, change of basis identities, and exact rational rank checks for K = 1 to 6. Propositions 2, 3, and 5 give invariance to a common conformer permutation, independent proper rigid motions, and consistent atom relabeling, each with a proof and a synthetic numerical check with errors around ten to the minus eight. The authors are explicit about what is not claimed, including nonlinear universality, an expressivity ordering relative to attention, and reflection invariance.

  3. The singleton centered residual is an elegant design. Re evaluating the same operator on the singleton input and subtracting gives an exact identity at K = 1 for every parameter setting, and the appendix example makes it concrete why masking the final diagonal is not equivalent. The ablation that removes the subtraction at identical parameter count shows a consistent degradation in mean score (ESOL and FreeSolv RMSE both up by 0.06, BACE ROC AUC down by 1.51 points), which supports the design within the evaluated setting.

  4. The empirical section is broad and the reporting is unusually candid. Six tasks, ten baselines reproduced locally under one random scaffold protocol, four interaction controls including an attention block matched to within 0.03 percent of the parameter count, a separately trained conformer cap study, and matched hardware profiling. The limitations section and Appendix G disclose the UniMol preprocessing mismatch, adapter defects, calibration risks, and release gaps rather than hiding them, which makes the paper easy to assess.

Critical weaknesses
  1. (strength of the main evidence): All results use three scaffold splits with a single initialization seed, no per split paired scores, and no significance tests. The margins over the strongest baseline are 0.02, 0.03, and 0.01 RMSE on ESOL, FreeSolv, and Lipo, while the reported standard deviations are 0.06, 0.42, and 0.06. UniMol v1 is within noise on all three regression tasks (0.78 vs 0.76, 1.73 vs 1.70, 0.57 vs 0.56) and its target scaler was fitted with validation targets while PaSTNet used training only standardization. PCNN is evaluated without contact edges and with 207K parameters against 668K, and DKO is a truncated diagonal variant with its own validation search budget. Appendix G also states that baseline conformer generation is model specific, so it is unclear whether the ensemble baselines (SPiCE, FACET, ConAN) saw the same conformer caches as PaSTNet. Could the authors provide paired per split scores, additional split seeds and several initialization seeds, a training only UniMol rerun, a contact enabled PCNN, and a statement of which conformers each baseline consumed? As it stands, best mean on five of six is not distinguishable from noise.

  2. (attribution and two puzzling orderings in Table 2): Most Table 2 differences are within one standard deviation (ESOL 0.76 ± 0.06 full vs 0.79 ± 0.08 attention; FreeSolv 1.70 ± 0.42 vs 1.76 ± 0.40 and 1.77 ± 0.42). Two patterns need an explanation. First, PaST without pair to pair context (0.85, 1.90, 82.02) is worse than removing PaST entirely (0.81, 1.82, 83.50) on all three tasks, so adding pair lifting and two tensor layers without contextual maps actively hurts. Is this an optimization effect, an interaction between centering and the purely local maps, or something else? Second, PaST without singleton subtraction is no better than the identity on ESOL (0.82 vs 0.81), which suggests that on ESOL nearly all of the gain is due to centering rather than relational context. Beyond that, the attention control changes geometry placement, pair state retention, and centering simultaneously, and the pair local control removes 49,152 parameters, so the central claim that explicit channel valued pair states matter is never isolated. A capacity matched pair local control and an attention control equipped with the same singleton subtraction would make the attribution far more convincing.

  3. (novelty relative to existing equivariant tensor layers): Theorem 1 is a direct restriction of the 15 element 2 to 2 basis of Maron et al. to the symmetric (9) and antisymmetric (2) subspaces, as the paper itself notes. Edge Transformers and Relational Attention already retain pair states, and Section 3.1 together with Appendix F.3 concede that pair aware attention can represent the same computation. What, then, is the new capability beyond applying a known 2 tensor equivariant layer independently at each path? In particular, would a plain IGN style layer with the full 15 map basis and no parity split perform the same? The symmetric and antisymmetric decomposition and the parity preserving products in Equation 9 are presented as design choices but are never ablated.

  4. (benchmark choice and physical scope): A conformer ensemble method is evaluated only on MoleculeNet tasks with small molecules, and Table 5 shows that roughly a quarter of ESOL and FreeSolv training molecules are singletons, on which PaST is exactly the identity. MARCEL, which the authors cite, and GEOM are the established ensemble benchmarks with larger flexible molecules and energy annotated ensembles; why are they not used? The conformers are selected by aligned heavy atom RMSD diversity from ETKDG and MMFF94s, energies never enter the model, and the pathwise pooling weights are not populations, so the physical meaning of the learned ensemble treatment is untested. The separately trained cap study also shows that going from K = 1 to K = 5 gains only 0.07 RMSE on ESOL, 0.08 on FreeSolv, and 0.55 points on BACE, with K = 3 below K = 1 on BACE, which raises the question of how much ensemble information these targets actually contain and whether the method is being evaluated where it should matter.

  5. (cost and scalability): Pair storage is O(P K squared d_r) per stage and explicit hydrogens inflate P, which itself grows as O(|V| times max degree cubed). At K = 10 PaSTNet uses 4.11 times the memory and 3.38 times the forward plus backward time of FACET, and the released code caps K at 5 and omits the K = 10, baseline, and interaction control runners. Given gains below 0.03 RMSE against UniMol and FACET, what is the argument that this cost is justified? How would the method handle drug like molecules with dozens of torsions or ensembles of thirty or more conformers, and is there any evidence that the gain grows with K or with molecular flexibility rather than saturating at K = 5?

  6. (known defects and symmetry gaps retained in the reported results): Appendix A.1 discloses that single and aromatic bonds occupy the same bond type slot and that the nominal length lookup always returns its default, leaving two descriptor channels constant, and that this adapter was kept for all reported results. Appendix A.5 discloses that the bond scale calibration includes diagonal zeros and has no positive floor. The predictor is not reflection invariant because of signed torsion harmonics, yet the splitter is chirality aware and BACE and ClinTox contain stereocenters; the synthetic check shows that reflection changes predictions by only about two times ten to the minus four, so chirality is effectively ignored in practice without being guaranteed either way. Have the authors rerun with a corrected bond adapter and a floored calibration, and does the absence of reflection invariance help, hurt, or not matter on these datasets? Finally, the reproducibility statement acknowledges that no run level manifest or paired split scores are provided and that the release cannot reproduce every table entry; can the full runners and manifest be released?

Questions for authors

Please refer to weaknesses section.

1 comment

Thank you for the authors' response. I have reviewed the rebuttal carefully. The authors have addressed the points I raised, and the clarifications improve my understanding of the submission. I'll keep my positive score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We sincerely thank all three reviewers for their careful reading, thoughtful evaluations, and constructive feedback. Their comments have been instrumental in strengthening the experimental evidence, clarifying the theoretical and architectural contributions, and improving the reproducibility of PaSTNet. We have substantially revised the manuscript, with key changes highlighted in blue. Below, we summarize the main improvements before addressing each reviewer's concerns and questions individually.

1. Stronger experimental reliability and baseline fairness (Tables 1–3; §4; Appendix G)

We expanded the evaluation to five random-scaffold split seeds, consistently across Tables 1–3. We also strengthened baseline implementations and comparability, including PCNN with nonbonded contact edges, Uni-Mol v1 with training-only target scaling, and DKO-Gated with the exact covariance spectrum. The revised Table 1 shows that PaSTNet achieves the best mean on four of six molecular property benchmarks, ranks second on the remaining two, and attains the lowest average within-task rank (1.33) among all eleven models, compared with 2.83 for Uni-Mol v1.

2. More rigorous attribution of PaST's contributions (Table 2; §4.3)

We expanded Table 2 to eight interaction operators, introducing stronger controls including capacity-matched pair-local PaST, singleton-centered attention, and a full 15-map equivariant tensor. Full PaST achieves the best mean across all three evaluated tasks, including improvements over the parameter-matched pair-local control. These comparisons provide more targeted evidence for the effectiveness of contextualized conformer-pair reasoning.

3. Clearer theoretical and architectural motivation (§§3.1–3.4; Appendix F)

We clarified the distinction between PaST and existing attention or equivariant tensor approaches, including an explicit mathematical example showing how one conformer-pair relation influences another before contraction. We also refined the scope of the branchwise linear completeness result and the singleton-centered residual, connecting the theoretical formulation more directly to the implemented architecture.

4. Expanded conformer-count and robustness evaluations (Tables 3, 8–9)

The updated five-seed conformer-cap study demonstrates meaningful improvements from K=1 to K=5, including a 9.5% reduction in ESOL RMSE and a 2.75-percentage-point increase in BACE ROC-AUC. We additionally evaluated fixed checkpoints under varying test-time conformer counts and examined sensitivity to different conformer-generation strategies.

5. Broader evaluation and computational scalability (Tables 4 and 10)

We added a conformer-sensitive evaluation on MARCEL Kraken-B5, where PaSTNet achieves 0.27 ± 0.03 Å MAE, compared with 0.31 ± 0.04 Å for SchNet under a shared five-split experimental protocol. We also extended matched-hardware resource profiling against FACET to K=30, explicitly quantifying the computational cost of relational pair states.

Finally, we expanded the experimental protocols, implementation details, symmetry analyses, and reproducibility documentation throughout the appendices.

We greatly appreciate the reviewers' valuable suggestions, which have helped make the manuscript more rigorous, comprehensive, and clearly motivated. We hope these substantial revisions address the central concerns and provide a stronger basis for evaluating PaSTNet's contributions. Detailed point-by-point responses follow.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.